MODEL BASED IMAGE RECONSTRUCTION USING DEEP LEARNED PRIORS (MODL).

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Model-based image reconstruction is crucial for enhancing image quality in medical applications.
• Deep learning, particularly Convolutional Neural Networks (CNNs), has shown promise in image reconstruction tasks.
• Existing methods often require large datasets and significant computational resources for training.

Purpose of the Study:

• To introduce a novel model-based image reconstruction framework utilizing a CNN-based regularization prior.
• To develop a recursive algorithm that integrates CNN denoising with data consistency enforcement.
• To create a deep network with shared CNN weights for improved efficiency and reduced data requirements.

Main Methods:

• A recursive algorithm alternating between CNN denoising and data consistency enforcement was employed.
• Unrolling the recursive algorithm resulted in a deep network trained via backpropagation.
• The framework incorporated a forward model and shared CNN weights across iterations for consistency and parameter reduction.

Main Results:

• The proposed framework demonstrated improved multichannel Magnetic Resonance Imaging (MRI) reconstructions.
• Shared CNN weights ensured consistency with the model-based formulation.
• Reduced trainable parameters led to lower training data requirements.

Conclusions:

• The developed framework offers a powerful approach for model-based image reconstruction.
• The integration of CNNs within a recursive, unrolled network architecture enhances reconstruction quality.
• This method presents a significant advancement over current state-of-the-art techniques in multichannel MRI.